Xueersi develops MathGPT, a large model for global mathematics enthusiasts

Recently, in the "ChatGPT War" that is being fought in full swing among major manufacturers, Xueersi has also joined in.

However, Xueersi has taken a different approach and chosen the self-developed large-scale mathematical model "MathGPT", which is based on problem-solving and lecture algorithms in the field of mathematics and has achieved phased results. .

Xueersi said that product-level applications based on this self-developed large model are expected to be launched within the year and will be targeted at mathematics enthusiasts and scientific research institutions around the world.

As the company’s core project, Xueersi started the corresponding team building, data, computing power preparation and technology research and development before the Spring Festival this year, and handed it over directly to CTO Tian Mi Responsible.

In addition, team building in Silicon Valley in the United States has also been started. It plans to establish an overseas algorithm and engineering team and recruit outstanding artificial intelligence experts from around the world.

The difference between MathGPT and large language model (LLM)

In March of this year, OpenAI officially released the large language model GPT-4. Subsequently, domestic Baidu and Alibaba also launched their own large model products.

However, the general language model is more like a "liberal arts student" and has excellent performance in tasks such as language translation, summarization, understanding and generation, but it is not good at solving, explaining, and There are obvious deficiencies in Q&A and recommendations -

"I often make mistakes when answering math problems. Although some math problems can be solved, the method is more adult-oriented and cannot target the knowledge structure of children of appropriate age. Adapt to the cognitive level."

In this regard, the leader of the Xueersi AI team said that this shortcoming is determined by the characteristics of the LLM model. The LLM large model comes from training on massive language texts, so it is best at language processing.

The industry tends to use LLM large models for reading and writing applications, but if you want to make a breakthrough in mathematical capabilities, you need to develop new large models.

Therefore, Xueersi is determined to form a team to specialize in MathGPT - a large model in the field of mathematics, using its many years of accumulation in mathematics and AI to target mathematics enthusiasts and Scientific research institutions should do a good job in basic mathematical work in the era of AI large models.

Xueersi hopes to make up for and overcome three problems of large language models through MathGPT:

First, the questions must be solved correctly. Now GPT results often have errors;

Secondly, the problem-solving steps must be stable and clear. Now the GPT problem-solving steps are different every time, and the generated content is often redundant;

Third, problem solving must be interesting and personalized. The current GPT explanations are too "academic" and mechanical, which is very unfriendly to children's learning experience.

Why do you do MathGPT? , is also the only member of the artificial intelligence "national team" in the education industry. It has many years of in-depth research in the field of artificial intelligence. As early as 2017, Xueersi established the AI ​​lab artificial intelligence laboratory.

According to public information, based on the help of the smart education artificial intelligence open innovation platform, Xueersi AI lab has won 16 championships and 6 runners-up in various top academic conference competitions; published in international journals and 31 high-level academic papers at conferences, including academic research in optical character recognition, images, natural language processing, speech and multi-modality, etc., and many papers were published in top computer vision conferences and top natural language conferences; It has applied for more than 220 patents, more than 150 authorized patents, and more than 60 software copyrights.

Xueersi AI lab’s awards in various top academic conference competitions

Xueersi, which "started with mathematics", has 20 years of mathematics teaching experience and has accumulated a huge amount of mathematics-related data. These data are necessary materials for MathGPT training.

In addition, Xueersi’s overseas business, Think Academy, is deeply loved by mathematics enthusiasts in several countries and regions around the world. Xueersi’s students perform well in international mathematics competitions such as IMO and AMC every year. Many students have won gold medals in the International Mathematical Olympiad.

Therefore, it is logical for Xueersi to choose to focus on MathGPT.

It is also understood that Xueersi Learning Machine will launch an "AI assistant" in the near future, covering composition assistant, speaking assistant, reading assistant, mathematics assistant and other related functions. The AI The product will start internal testing on May 11.

Challenges and technical problems of MathGPT

How to use large language models to serve all walks of life is a focus issue in current society.

For example, in the field of education, Duolingo, Quizlet, Khan Academy and other products mainly cooperate with OpenAI to make fine-tuning and interface calls on the GPT large model to enhance the original product experience.

But there are also some fields such as mathematics, medicine, etc., which require AI to be accurate, clear, have strong logical reasoning capabilities, and have a low fault tolerance rate. The current performance of general LLM is still poor. Breakthroughs in the above areas have not been achieved, and it is unclear whether breakthroughs are possible in the future.

Taking the field of mathematics as an example, there are several main schools on the market.

For example, products such as Photomath, Microsoft Mathematics, Mathway, and WolframAlpha that focus on mathematical calculations acquired by Google mainly use non-LLM traditional AI technology and database methods to solve mathematical problems.

Companies taking the AGI route are trying to make general LLM "more mathematical". For example, GPT-4 performs better on mathematical tasks than the previous version 3.5, and Google's Minerva model It is also specifically tuned for mathematical problems.

Xueersi has chosen another path less traveled. It does not make fine-tuning and interface calls based on existing LLM, nor does it make general-purpose LLM. Instead, it develops its own " "Mathematical Big Model" MathGPT is committed to creating independent, stable, sustainable and high-quality learning solutions.

Under the wave of continuous evolution of large language models, the advantages and disadvantages of different technical route choices still need to be discussed and verified.

Whether Xueersi’s self-developed independent large-scale MathGPT model is established, whether it can surpass the performance of general models on mathematical tasks, and whether it can better match the mathematical learning scenarios of different groups of people, this question remains. The answer needs to be found in innovative practice.

With the deepening development of the entire industry and more and more talents participating in this field, I believe that more mature solutions will be seen in the near future.

The above is the detailed content of Xueersi develops MathGPT, a large model for global mathematics enthusiasts. For more information, please follow other related articles on the PHP Chinese website!